Reidentification by Relative Distance Comparison
read more
Citations
DeepReID: Deep Filter Pairing Neural Network for Person Re-identification
Person re-identification by Local Maximal Occurrence representation and metric learning
Beyond Part Models: Person Retrieval with Refined Part Pooling (and A Strong Convolutional Baseline)
Harmonious Attention Network for Person Re-identification
Person Re-identification: Past, Present and Future
References
Distance Metric Learning for Large Margin Nearest Neighbor Classification
Distance Metric Learning with Application to Clustering with Side-Information
Information-theoretic metric learning
An efficient boosting algorithm for combining preferences
An Efficient Boosting Algorithm for Combining Preferences
Related Papers (5)
Viewpoint Invariant Pedestrian Recognition with an Ensemble of Localized Features
Frequently Asked Questions (16)
Q2. What are the future works in "Re-identification by relative distance comparison" ?
Nevertheless, it will be interesting to integrate an explicit background segmentation step into the proposed framework in the future.
Q3. How many features were extracted for each stripe?
For each stripe, the RGB, YCbCr, HSV color features and two types of texture features extracted by Schmid and Gabor filters were computed across different radiuses and scales, and totally 13 Schimid filters and 8 Gabor filters were obtained.
Q4. Why are the relative distance comparison formulations in these works not quantified?
the relative distance comparison formulations in these works are not quantified using logistic function for soft measure, and crucially they are used as an optimisation constraint rather than an objective function.
Q5. How many feature channels were constructed for each stripe?
In total 29 feature channels were constructed for each stripe and each feature channel was represented by a 16 dimensional histogram vector.
Q6. How many people were captured in the i-LIDS MCTS dataset?
In the i-LIDS MCTS dataset, which was captured indoor at a busy airport arrival hall, there are 119 people with a total 476 person images captured by multiple non-overlapping cameras with an average of 4 images for each person.
Q7. What is the reason why RDC is better than other methods?
The better performance of RDC is mainly due to the logistic function based modelling that enforces a softer constraint on relative distance comparison and exploiting second-order rather than first-order feature quantification.
Q8. How long did it take to learn RDC?
In their experiments, for VIPeR with p = 316, it took around 15 minutes for an Intel dual-core 2.93GHz CPU and 48GB RAM server to learn RDC for each trial.
Q9. What is the main reason for the inferior performance of the compared alternative learning approaches?
Overall the results suggest that over-fitting to under-sampled training data is the main reason for the inferior performance of the compared alternative learning approaches.
Q10. How many people were captured in the VIPeR dataset?
The VIPeR dataset6 is a person reidentification dataset available consisting of 632 people captured outdoor with two images for each person with normalised size at 128 × 64 pixels.
Q11. What is the reason why MCC gives the performance to RDC when the training set is?
It is noted that, benefiting from being a Bayesian modelling, MCC gives the most comparable results to RDC when the training set is large.
Q12. How can one reduce the number of potential matches?
By modelling the transition time between two camera views one can reduce the number of potential matches while also using the probability distribution of transition time as a feature [12], [25], [24], [22].
Q13. What is the criterion for the iteration of the algorithm?
The iteration of the algorithm (for > 1) is terminated when the following criterion is met:r (w ,O )− r +1(w +1,O +1) < ε, (12)where ε is a small tolerance value set to 10−6 in this work.
Q14. What is the significance of the logistic function for learning a person re-identification model?
The results show that without the logistic modellingfor differentiating the margin in the difference information from different types, the RDC-MMC model performs much worse for person re-identification.
Q15. Why is the ensemble RDC better than the batch model?
The better performance of ensemble RDC is likely due to the fact that the ensemble learning process can effectively alleviate the local optimum of the iterative algorithm for optimising RDC.
Q16. What is the space complexity of the ensemble learning process?
After generating the weak RDCs, the ensemble learning process itself has a space complexity of O(H · (( 1L − 1 L2) · N3 + ( 1L − 1) · N2)), where H is the number of groups (i.e. the total number of weak RDC models).